Windows: FFmpeg `libvmaf` filter creates a mojibake filename for non-ASCII `log_path`

[AntaresFeng]

Description

On Windows, the log_path option of FFmpeg's libvmaf filter may create a file with an incorrectly encoded filename when the requested filename contains non-ASCII characters.

For example, requesting 你好.json (means “Hello” in Chinese) creates 浣犲ソ.json (mojibake) instead. The encoding mismatch can be demonstrated directly:

>>> "你好".encode("utf-8").decode("cp936")
'浣犲ソ'

The problem affects the filename only. The generated file is a valid VMAF JSON log, and its contents are not garbled or corrupted.

Changing the console code page to 65001 does not help.

Reproduction

ffmpeg -hide_banner -f lavfi -i "testsrc2=size=1280x720:rate=30:duration=1" -filter_complex "[0:v]split=2[dist][ref];[dist][ref]libvmaf=log_fmt=json:log_path='你好.json'" -f null -

After the command completes, the current directory contains a valid JSON log named: 浣犲ソ.json rather than the requested: 你好.json

Other Cases:

测试.json (means "test") -> 娴嬭瘯.json
热力图.json (means "heatmap") -> stderr: could not open file: 鐑姏鍥?json

Comparison with FFmpeg's psnr filter

FFmpeg's built-in psnr filter does not exhibit this problem with the same non-ASCII filename:

ffmpeg -hide_banner -f lavfi -i "testsrc2=size=1280x720:rate=30:duration=1" -filter_complex "[0:v]split=2[dist][ref];[dist][ref]psnr=stats_file='你好.log'" -f null -

This command creates the correctly named file 你好.log

The psnr filter opens its output through FFmpeg's UTF-8-aware file-opening support, whereas the libvmaf output file is opened inside libvmaf. This comparison also shows that the problem is not caused by the shell, the console code page, or FFmpeg's filtergraph parsing of the filename.

Environment

• Windows 10 with default console code page 936
• FFmpeg n8.1 build by BtbN, with libvmaf statically linked into avfilter-11.dll
• Reproduced with libvmaf commit 0f9912e

Root cause analysis

FFmpeg passes the log_path string to libvmaf, which creates the output file in vmaf_write_output(). The current implementation uses a narrow fopen() call:

int vmaf_write_output(VmafContext *vmaf,
                      const char *output_path,
                      enum VmafOutputFormat fmt)
{
    FILE *outfile = fopen(output_path, "w");
    ...
}

On Windows, narrow CRT file APIs do not reliably interpret a UTF-8 pathname when the active ANSI code page is CP936. This explains why the UTF-8 bytes for 你好 are interpreted as CP936 and produce the filename 浣犲ソ.

A possible fix would be to use a UTF-8-aware file-opening path on Windows, for example by converting the UTF-8 pathname to UTF-16 and calling _wfopen(). A Windows regression test should verify the exact filename created for a non-ASCII log_path.